View
4
Download
0
Category
Preview:
Citation preview
1
Who wins and who loses from state subsidies?
Jun Du
Aston Business School
Sourafel Girma
University of Nottingham
Holger Görg
Kiel Centre for Globalization at the Kiel Institute for the World Economy and the University
of Kiel
Ignat Stepanok
Institute for Employment Research (IAB), Nuremberg
Abstract
China is perceived to rely on subsidizing firms in targeted industries to improve their
performance and stay competitive. We implement an approach that allows for the joint
estimation of direct and indirect effects of subsidies on subsidized and non-subsidized firms.
We find that firms that receive subsidies experience a boost for productivity. However, our
approach highlights the importance of indirect effects, which are generally neglected in the
literature. We find that, in general but not always, non-subsidized firms experience reductions
in their productivity growth if they operate in a cluster where other firms are subsidized.
These negative externalities depend on the share of firms that receive subsidies in the cluster.
We interpret our results in the light of a simple heterogenous firm type model, which
highlights the implications of subsidization, in a competitive environment of firms, may
potentially harm non-subsidized firms.
JEL Classification: H25, H32, L25
Keywords: Subsidies, Firm performance, Treatment effects, Externalities, China
2
1. Introduction
In recent years, China has run a trade surplus with nearly 80% of the countries in the
world.1 The most striking trade imbalance between the US and China has stirred up the recent
trade war between the two countries, giving rise to uncertainties for the global economy and
likely losses for both sides. As frequently expressed in public, one of the reasons behind the
US trade war against China is the perceived unfair competition in its trade practices induced
by the prevailing state subsidies.
Indeed, China is frequently noted for its policy of subsidizing firms in targeted
industries. Some recent examples include electric cars, steel, and solar panels, all of which
were discussed as controversial in the media.2 Underlying this policy is, presumably, the
assumption on the part of Chinese policy makers that subsidies help Chinese firms improve
performance and thus competitiveness. Yet, neither the theoretical nor empirical base for
such an assumption is clear-cut. This paper provides fresh evidence on this issue using
Chinese firm-level data. We estimate the effect of subsidization on firm level productivity,
paying particular attention to the fact that, while subsidized firms may benefit, non-
subsidized firms may be harmed by such a policy as an unintended consequence.
Economic theory offers contrasting views regarding the merits of firm- or sector-
specific subsidies as tools of industrial development policy (see Haley and Haley, 2013a, and
references therein). On the one hand, subsidies are seen as an effective way of overcoming
market failures such as informational asymmetries in credit markets, or for protection of
infant industries, and hence as useful policy instruments for improving firm performance and
creating comparative national advantage. This view would thus be in line with the assumption
that subsidies give a benefit to Chinese firms. On the other hand, however, subsidies are
1 According to the WTO’s Integrated Data Base (IDB), in 2016, China ran trade surplus with 169 countries
among 215 countries that have statistics. 2 See, for example, the Financial Times at http://www.ft.com/cms/s/0/a55e7d36-db8a-11e5-a72f-
1e7744c66818.html#axzz4BeAlJ6Ax and http://www.ft.com/intl/cms/s/0/6d4e6408-5e68-11e2-b3cb-
00144feab49a.html#axzz4BeAlJ6Ax (accessed 15 June 2016), and Haley and Haley (2013b).
3
perceived as distorting the efficient allocation of resources in the economy, which may lead
to worsening firm performance and an aggregate welfare loss (Mueller, 2003).
There are several empirical papers that seek to evaluate the direct impact of public
subventions on the subsidy-receiving firms’ performance in a number of countries (see inter
alia Bernini and Pellegrini, 2011; Cerqua and Pellegrini, 2014; Howell, 2017; Girma et al.,
2009; Görg et al., 2008). However, subsidies, importantly, have a broader impact than just
the direct impact on recipients alone. Subsidies may inflict positive or negative externalities
(spillovers) on non-subsidized firms. Externalities may arise as subsidized firms change the
competitive environment, which are likely to affect the strategy, conduct and performance of
non-subsidized firms. This means that in addition to estimating the direct impact of subsidies
on subsidized firms, for a comprehensive appraisal it is important to evaluate the indirect
effects on non-subsidized firms.
However, most empirical studies (including those cited above) on the effect of
subsidies do not allow for spillovers to firms that do not receive subsidies. Rather, they focus
on estimating the direct effect of the subsidy on the receiving firm. A notable exception is De
Mel et al. (2008) who look at the impact on returns to capital of small cash grants to
microenterprises in Sri Lanka.3 Another issue that is generally not considered in the literature
is that the strength of externalities generated by subsidization may in turn impact on the
relative magnitudes of the direct or indirect effects of the subsidy. For example, the more
subsidized firms we have in a particular geographic area, the lower may be the gain for the
treated firm from receiving a subsidy, or the stronger may be the negative externality on the
non-subsidies firms.4
3 They adopt a field experiment approach where a small number of grant recipient firms were randomly chosen
to avoid the problem of selectivity in subsidy receipt, before going on to estimating the direct effects of
receiving the subsidy by comparing recipient and control group firms. Spillovers are taken into account by
controlling for the number of treated firms within a limited geographic radius (i.e. a co-location). 4 The existence of indirect spillover effects, or the dependence of the effects on the strength of the externality,
are usually ruled out in the microeconometric evaluation approaches commonly employed in the literature by
4
The novel estimation approach employed in this paper allows us to tackle this issue,
and hence address an important gap in the literature that debates the role of the state through
subsidizing businesses. We exploit observational data on subsidy receipt from a large-scale
firm level panel dataset for the Chinese manufacturing sector. We adapt the methodology
developed by Girma et al. (2015), which enables us to estimate differing direct or indirect
effects depending (possibly non-linearly) on the level of subsidizationsubsidization in a town-
cluster. Importantly, our identification strategy recognizes that there are two levels of
selection. The first issue, as is well recognized in other studies, is that the selection of
subsidized firms is unlikely to be random. This is why studies use micro-econometric
evaluation techniques, such as propensity score matching. However, there is a second
selection problem that previous work has overlooked. When modelling the levels of
interference, say, as the number or share of treated firms within a province or industry, the
distribution of treated firms across such clusters is also unlikely to be random. There may, for
example, be deliberate government policy towards attracting certain types of firms to certain
provinces or sectors, which also provide subsidies. Or there may be other non-random factors
determining the location of subsidized firms. This selection problem has generally not been
recognized in the literature thus far. The approach implemented here allows us to deal with
both selection problems using generalized propensity score matching techniques at the two
levels. Furthermore, given the particular focus on Chinese state-owned enterprises (SOEs) in
the subsidy debate, we estimate performance effects separately for private owned firms,
SOEs and foreign owned firms.
assumption: they implicitly assume no interactions between firms; or in the parlance of the econometric
literature, the Stable Unit Treatment Value Assumption (SUTVA). This assumption essentially posits that an
individual outcome does not depend on the treatment status of others. Hence, the estimation of the spillover
effects described above are ruled out by assumption – an assumption that is unlikely to hold in practice, given
the very plausible arguments and arising evidence as to why one may expect subsidies to have externalities on
non-subsidized firms.
5
Our empirical analysis identifies a positive direct effect: Chinese state subsidies have
evidently benefited recipients by enhancing their productivity, irrespective of their
ownership. Importantly, the magnitude of this positive direct effect depends on how widely
subsidies were given out. However, the direct effects of subsidies do not tell the whole story.
The estimation of indirect effects of subsidies reveals mainly negative effects on
unsubsidized firms. Aggregating direct and indirect effects into a (weighted) total effect
shows that this negative indirect effect tends to dominate. For all three firm types, subsidies
have an overall negative effect, especially in clusters with fairly high shares of subsidized
firms.
In order to make sense of our empirical results, we build a simple theoretical model
with heterogenous firms, which fits most of the observed patterns. In a related paper, Pflüger
and Suedekum (2013) also have a heterogeneous firm model and show that giving subsidies
to firms increases competitive pressures and allows in equilibrium only more productive
firms to enter a market, thus leading to higher average productivity of operating firms. Our
empirical results however show a more complicated picture, where spillover effects from
subsidies can lead to deteriorating average productivity of firms. We propose some simple
extensions of the heterogeneous firm model that can generate those results and provide a
good fit between theory and data. Specifically, our model shows how the direct effect on
subsidized firms but also the spillover effect on non-subsidized firms depend on how many
firms receive subsidies in the clusters. Subsidies change the competitive environment, so the
total amount of subsidies received affects firm selection into entering a market. That way
subsidies affect average productivity in a cluster, without making any additional assumptions
about individual firm investment in, say, productivity or innovation.
The rest of the paper is structured as follows. The next section discusses some of the
institutional background to the policy of using subsidies in China. This is followed by a brief
6
description of our data set in Section 3. Section 4 sets out the econometric methodology, and
Section 5 presents and discusses empirical results. Section 6 then discusses a theoretical
framework that fits many aspects of the empirical results. Section 7 concludes.
2. Institutional background and literature
The institutional foundation that governs China’s economic dynamics can be described as a
regionally decentralized authoritarian system in which the central government incentivizes
local officials to promote regional economic growth (see Xu, 2011 and references therein).
This has resulted in regional economic decentralization where local governments actively
engage in shaping the business and economic landscape of their respective regions, and
directly intervene in relation to businesses’ investment and operational decisions.
A distinguishing feature of Chinese state capitalism is the use of capital controls by
the government, including soft budget constraints (Kornai et al., 2003), influencing local
banks (Lin et al 2008), or offering easy access to land and other economic resources to
politically protected firms (Du and Girma, 2010). Perhaps one of the most controversial
practices is the use of outright subsidies.
While government subsidies are by no means exclusive to China, the reason that the
latter’s case is attracting such interest stems from the fact that subsidies are spread over a
large spectrum of firms and a broad range of sectors (Haley and Haley, 2013a). Shao and Bao
(2011) find that over the years 2000-2006, about 13-19% of all firms reported in the
Industrial Census receive subsidies and the percentage has been increasing over time. Girma
et al. (2009) report that over the period 1998 to 2004, government production subsidies to
manufacturing firms amounted to more than $100 billion. In a more recent study, Haley and
Haley (2013a) combine official statistics with information from industry analysts and policy
documents, and they estimate that China may have spent well over $300 billion on its largest
SOEs between 1985 and 2005.
7
As in Howell (2017) and Girma et al. (2009), we use data on production-related
subsidies that are allocated to firms. There are generally several reasons why governments
subsidize enterprises: industrial development, export promotion, supporting firms to innovate
and securing a national advantage in leading industries (WTO, 2006). An additional specific
motivation for the Chinese government to subsidize SOEs is to avoid a worsening of
unemployment rates and social riots due to possible bankruptcies of SOEs (Luo and
Golembiewski, 1996).
Having experienced a prolonged economic boom, China now faces the growing
concerns of unsustainable high investment rates and soaring production costs. Despite
increasing R&D spending, the country’s rate of productivity growth remains relatively low,
and China appears to be heading towards the “Middle Income Trap” (Woo et al., 2012).
Recent evidence suggests that resource misallocation problems both within and between
firms, can partly explain the slow aggregate productivity performance (Hsieh and Klenow,
2009; Du et al., 2014). As Hsieh and Klenow (2009) point out, government policies may
well have a role to play to account for such misallocation.
Aghion et al. (2015) look at the implications of subsidies (as one aspect of industrial
policy) for firm level productivity in China. Using similar data to ours, they find that
industry-city combinations where the correlation between subsidy receipt and the level of
competition is higher also have firms with higher productivity growth. Also, a greater
dispersion of subsidies within an industry-city combination is associated with higher firm
level productivity growth. They control for subsidy receipt at the level of the individual firm,
which is positively correlated with productivity growth. While our results are not strictly
comparable given the different approach used, our findings on the direct effects are in line
with theirs. However, they do not look at indirect effects, nor do they allow the effects to vary
with the strength of subsidization. In contrast to Aghion et al. (2015), Howell (2017) finds
8
negative direct effects when looking at the relationship between subsidies and productivity
growth for Chinese firm level data using propensity score matching. However, he also does
not allow for externalities, which may likely bias results.
We therefore investigate the link between government intervention through subsidies and
firm productivity, considering both the direct impact of receiving a subsidy on the firm’s own
performance as well as the indirect spillover effects on other firms.
3. Data and exploratory analysis
3.1 Data
We draw on firm level data from the Chinese manufacturing industry. The dataset is
based on the Annual Reports of Industrial Enterprise Statistics, compiled by the China
National Bureau of Statistics. The enterprises covered by this dataset account for an estimated
85–90 percent of total output in most industries. For the purpose of this analysis, we have
more than 300,000 firms over the period 1998-2007. The precise definition of the variables
used in the analysis is given in Table 1.
[Table 1 about here]
Table 2 shows the value of total subsidies and the number of subsidized firms by
ownership category, distinguishing private firms, foreign invested firms and state-owned
enterprises (SOEs). It is noticeable that all categories of firms received substantial amounts
of subsidies. For example, in 2007, more than 23 Billion USD was paid to 25,673 private
firms, and just 4,077 state-owned enterprises (SOEs) received 6.7 Billion USD worth of
production subsidies. Figure 1 reveals that the proportion of firms receiving subsidies has
been increasing steadily from 1998 to the middle of the 2000s. Also, time series plots of the
average amount of subsidy amongst subsidized firms given in Figure 2 show that, not
surprisingly, SOEs enjoyed the largest number of subventions over the study period.
[Table 2 about here]
9
Table 3 gives some summary statistics on variables of interest by subsidy status. The
most noteworthy difference is that firms that are subsidized in year t are about 10 times more
likely to have received a subsidy also in t-1 and t-2. This suggests that subsidy receipt tends
to be path-dependent.
[Table 3 about here]
3.2 Selection into subsidy
Given the above statistical observations, in order to better understand the pattern of subsidy
distribution, we estimate the determinants of subsidy receipt amongst Chinese firms over the
period 1998-2007, conditional on variables that are all measured in the period prior to
subsidy receipt. Drawing on the literature discussed in the previous sections, the regression
model includes the following pre-subsidy firm level characteristics: past history of subsidy
receipt, firm size (level of employment), TFP, TFP growth trend, firm age, existence of
political connections, debt (a proxy for access to formal financing channels), history of loss-
making, and ownership type (private being the baseline group). These variables capture the
selective nature of subsidy receipt (see Table 1 for details of variable definitions).
Moreover, taking into account the regionally decentralized nature of China’s policy
making milieu, we include a second group of conditioning variables that are the cluster-level
averages of the firm-level variables, excluding the firm’s own value. For the purpose of our
empirical implementation geographic clusters are based on three-digit administrative division
codes which roughly identify prefectures. The manufacturing enterprises in our dataset are
located in 74 such prefectures which we henceforth simply refer to as town-clusters. These
are designed to capture the spatial dependence amongst firms given that subsidies in China
are largely administered by local government authorities. The model is estimated using a
spatial logistic regression which also includes time, ownership and industry dummies.
10
The log odds ratios from the logistic regression are reported in Table 4. The strongest
predictor of subsidy both at the firm and spatial levels is subsidy receipt in the past, as also
suggested in the simple summary statistics in Table 3. Interestingly, young, highly productive
and larger firms are more likely to receive subsidies all else equal, as are loss-making and
politically connected ones. Moreover, state-owned enterprises are significantly more likely to
receive subsidies than private or foreign firms. Overall the regression results show that the
decision to allocate subsidies amongst firms is not a random process, but rather one that is
systematically correlated with firm and cluster level variables.
[Table 4 about here]
This observation motivates our empirical strategy which we discuss in detail below.
4. Estimating the effects of subsidy
The aim of the empirical exercise is to estimate direct and indirect treatment effects of
subsidy receipt on productivity. These may potentially differ depending on the level of
subsidization in a cluster. As noted in the descriptive analysis in Section 3, firm’s selection
into subsidy is unlikely to be random, and neither is the selection of town clusters. This
motivates the evaluation approach adopted in this study. Hence, controlling for selection at
firm and cluster level is essential. In this section, we set out our basic identification strategy.
In order to deal with the two levels of selection, our empirical approach proceeds in
two steps (see Girma et al. 2015 for a more detailed description). To tackle selection at the
firm level, we firstly estimate the firm-level relationship between subsidy receipt and
productivity separately for each town-cluster, using data at the firm level. In a second step, to
allow for selection at the cluster level, we take the share of subsidized firms in a cluster as
treatment, using data at the cluster level. In both steps, we apply generalized propensity score
matching techniques.
11
First step estimation
In the first step we, start by estimating the relationship between subsidy receipt and
productivity separately for each of the 74 town-clusters and 3 firm types (private, SOE,
foreign). We define a binary treatment variable dirt = 1 if firm i in cluster r receives a subsidy
in year t, and dirt = 0 if not. This treatment variable is then used as independent variable in a
productivity regression using the firm level data for a given cluster and firm type. In order to
take into account selection at the firm level, we estimate the outcome equation using inverse
propensity-score weighted regression and controlling for the pre-treatment covariates (Bang
and Robins, 2005, Hirano et al., 2003). Note that the outcome variable, firm-level
productivity is defined as the change relative to t-1, akin to using a difference-in-differences
strategy combined with propensity score matching.5
For each cluster and firm type, this implies that we firstly generate the firm-specific
propensity-scores of being treated via a logistic regression with a rich list of pre-treatment
covariates subject to balancing conditions being satisfied. The list and precise definition of
the pre-treatment covariates can be found in Table 1, and covariate balancing tests results are
reported in Appendix Tables A2-A4.
Using the estimated propensity scores we then estimate the following outcome
equation (after imposing the common support condition) for each cluster and firm type
separately via inverse probability weighted regression,6
𝑦𝑖𝑟 = 𝛼 + 𝛽𝑑𝑖𝑟 + 𝐹(𝑋; 𝛿 ) + 𝑒𝑟𝑟𝑜𝑟; i=1…N. (1)
5 This estimation strategy provides two opportunities to adjust for selection on observables by combining
inverse probability reweighting with regression covariates adjustment. The identifying assumption is selection
on observables. To the extent that there are unobservables that are correlated with both the treatment conditional
on observables and the change in productivity, our results would potentially be biased. 6 Treated firms receive a weight of 1/ and non-treated firms a weight of 1/(1-
12
where y is the change in firm level productivity and F(.) is a function of the pre-treatment
covariates vector X. From these regressions we can then calculate the cluster specific
average potential outcomes for each firm type,
�̅�𝑟1 =
1
𝑁∑ �̂�𝑁
𝑖=1 + �̂� + 𝐹(𝑋; 𝛿 ) 𝑎𝑛𝑑 �̅�𝑟0 =
1
𝑁∑ �̂�𝑁
𝑖=1 + 𝐹(𝑋; 𝛿 ) (2)
Second step estimation
In the second step, the cluster and firm type level average potential outcomes, �̅�𝑟1 and �̅�𝑟
0
estimated in the first step are treated as the "outcome" variables. The proportion of subsidized
firms in the cluster, 𝑠𝑟 is taken to be the continuous "treatment" variable. Since the treatment
dosage 𝑠𝑟 is again unlikely to be randomly distributed (e.g. due to endogenous difference
between local governments when it comes to the extent of subsidy usage), we employ the
causal inference approach for continuous treatments (Hirano and Imbens, 2004; Imai and van
Dye, 2004). A key result from this literature is that causal inference can be conducted by
conditioning on the generalized propensity score (GPS), which is nothing but the conditional
density of the treatment given some pre-treatment balancing covariates.
It is clear that our treatment dosage variable 𝑠𝑟 is continuous and bounded between 0
and 1. Accordingly we generate the GPS conditional on pre-treatment cluster level covariates
using the fractional logit model due to Papke and Wooldridge (1996). In the empirical
implementation, the vector of observable pre-treatment characteristics consists of cluster-
specific averages of the firm level covariates discussed in section 3.2 above.
Defining Z to be the vector of cluster level pre-treatment covariates; �̂� the vector of
estimated coefficients from the fractional logit model and 𝜔𝑖 ≡ 𝑍𝑖′�̂�, for a given level of
treatment intensity 𝑠, the GPS conditional on Z can be obtained as
�̂�𝑟 = [𝑒𝜔𝑖
1+𝑒𝜔𝑖]
𝑠𝑟
[1
1+𝑒𝜔𝑖]
1−𝑠𝑟
(3)
13
The expected values of each of the two cluster and firm type level potential outcomes
( �̅�𝑟𝑑, d=0, 1) conditional on �̂�𝑟 and 𝑠𝑟 can then be obtained using quadratic approximation
(Hirano and Imbens, 2004) as:
𝐸[𝑦𝑟𝑑|�̂�𝑟 , 𝑠𝑟 ] = 𝛽0 + 𝛽1�̂�𝑟 + 𝛽2𝑠𝑟+𝛽3�̂�𝑟
2+ 𝛽4𝑠𝑟
2 +𝛽5�̂�𝑟𝑠𝑟 (4)
The above polynomial regression is based on r (number of clusters that are on the
common support of the GPS) observations, and the sample average potential outcomes are
obtained as
�̅�𝑟𝑑 =
1
𝑅∑ �̂�0 + �̂�1�̂�𝑟 + �̂�2𝑠𝑟 + �̂�3�̂�𝑟
2+ �̂�4𝑠𝑟
2 + �̂�5�̂�𝑟𝑠𝑟𝑅𝑟=1 (5)
Subsequently we calculate the predicted values �̅�𝑟𝑑 for the two firm level treatments d
and the continuous cluster-level treatment s.
Calculating treatment effects
Using these predicted values as potential outcomes, we can then calculate treatment
effects, using insights from the recent statistical literature (e.g. Hudgens and Halloran, 2008).
Firstly, we can calculate a direct treatment effect �̅�𝑠𝑠 10 = �̅�𝑠
1 − �̅�𝑠0 as the difference in
productivity between subsidized (1) and non-subsidized (0) firms for a given level of the
proportion of subsidized firms s in a cluster-firm type group.
Secondly, the indirect treatment effect is defined as �̅�𝑠0 00 = �̅�𝑠
0 − �̅�00 , hence, the
difference in productivity between non-subsidized firms (0) in a cluster with proportion of
subsidized firms s and in a cluster without any subsidies.
Based on these two treatment effects we calculate an overall or total treatment effect,
described in more detail below, as a weighted sum of the direct and indirect effects.
5. Findings
14
We firstly estimate the direct and indirect effects separately for any level of
subsidization s in a cluster as described above. Given that s is between 0 and 100 per cent, the
presentation of all treatment effects in a single table is not practical. Instead, we plot all
estimated direct and indirect effects among private, state-owned and foreign-owned firms
along with their 95% confidence intervals in Figures 3 and 4.7 An immediate observation
from these plots is that the share of subsidized firms in a cluster matters significantly for the
magnitude of both the direct and indirect effects among all types of firms, both statistically
and economically.8
Direct effects: the more, the merrier
We first focus on the direct effects of subsidies. As shown in Figure 3, there is a
positive direct effect for all levels of s for all firms. This suggests that subsidized firms have
higher productivity as a result of receiving a subsidy, which is in line with the idea that
subsidies reduce a recipient’s marginal cost of production.
Importantly, the strength of the direct effect depends on the level of subsidization in a
cluster, and the relationship between the productivity-enhancing effects and the proportion of
firms that receive subsidies is nonlinear. Our estimates suggest a positive direct effect of
subsidies that for the most part increases in s.
The relationship between s and the productivity effect appears fairly similar for
private domestic and state-owned firms. For these firms, in a cluster with 5 percent share of
subsidized firms, firms that receive a subsidy on average enjoy a productivity increase by 2
percent. In a cluster with 50 percent share of subsidized firms, the direct effect is stronger at
around 6 percent for private firms and 8 percent for SOEs, respectively. The direct effect,
7 The calculation of standard errors is complicated by the fact that the potential outcomes are estimated. Hence,
we compute bootstrapped standard errors via resampling with replacement, see Girma et al. (2015) for more
details. 8 Note that s in all graphs is calculated for all firms (private, SOE, foreign), so s reflects in all cases the share of
all subsidized firms relative to all firms in a cluster.
15
while always positive, firstly decreases in s and then turns to increasing in s from around s =
17 percent.
It is noticeable that the graph for foreign firms is quite different. While these are
generally privately owned, they are different from domestic firms in that they are affiliates of
foreign-owned multinationals and as such may be expected to behave differently than
domestic firms (e.g., Bellak, 2004). There seems to be only limited additional productivity
gain for foreign firms from securing government subsidies for levels of s lower than about 75
percent. This is not surprising, given that foreign invested firms in China are mostly resource-
rich and likely to have received other forms of preferential treatments such as tax reductions
or exemptions of utility bills (Klitgaard and Rasmussen, 1983).
Indirect effects: the unintended losers of subsidies
Turning to the indirect effects of subsidy-giving, it is clear that the signs depend on s,
the proportion of firms that are subsidized in a town-cluster, and firm ownership. As shown
in Figure 4, unsubsidized SOEs also experience productivity-reductions due to the state
subsidies. The negative effect deteriorates initially but then improves until four-fifths of all
firms are subsidized. Then, the indirect effects turn positive towards the end of the
distribution.
A similar picture emerges for foreign-owned firms. The indirect effects of subsidies
for foreign firms are negative initially, suggesting that subsidies harm unsubsidized firms’
productivity. In fact, there is a worsening productivity-reducing effect for unsubsidized
foreign firms as more subsidized firms populate, and then that starts to improve as more than
20% of all firms are subsidized. With higher levels of subsidization in a cluster, these turn
less negative and eventually positive when 45% or more of all firms are subsidized, similar to
the case of indirect effects for SOEs.
16
However, a very different picture emerges with domestic private firms. The indirect
effect on unsubsidized private firms starts being positive. However, this decreases sharply
with s, until the level of the coverage reaches around 25% of subsidized firms, when the
indirect effect of subsidies becomes negative.
Overall effects
The presence of both winners and losers of the state subsidies makes the overall
economic impact and interpretation ambiguous. Next, we adopt some “back-of-the-envelope”
calculations on the overall effect of subsidies on productivity among treated and non-treated
firms. To do this, we follow the standard approach to define a “total treatment effect” as the
sum of the direct and indirect treatment effect (Hudgens and Halloran, 2008). As shown
above, the strength of the overall effect depends on the relative number of subsidized and
non-subsidized firms, i.e., what share of firms experiences the direct and how many the
indirect effect. Hence, we calculate an overall effect as a weighted average of the direct and
indirect effect, weighted by the relative share of the two groups of firms. These calculations
are included in the appendix.
Keeping in mind that the direct effect is always positive, the overall effect will
certainly be positive as long as the indirect effect is also positive. For private firms, as we can
see from Figure 3, this is the case up to a level of s = 25 percent. After this point, we
calculate that the overall effect as a weighted sum of direct and indirect effect is always
negative. For example, at s = 30 percent (where 70% of firms do not receive a subsidy), the
overall effect is 0.018 × 0.3 + −0.016 × 0.7 = −0.0058. At s = 50 percent (i.e., the
groups of subsidized and non-subsidized firms are equally large), the direct effect is around
0.07, while the indirect effect is around – 0.12. Hence, the overall negative spillover effect
outweighs the positive direct effect. In sum, some subsidies benefit private firms irrespective
of whether they are subsidized or not, but too much subsidization hurts them.
17
For SOEs and foreign firms, the overall negative spillover effects on unsubsidized
firms outweigh the positive direct effects on subsidized firms, for as long as no more than 45-
50% of all firms are subsidized. That gives a largely negative overall weighted total effect for
SOEs and foreign firms. Table A4 summarises the calculation.
In our data, the mean value of subsidized firms in a cluster is 15 percent, while the
75th
percentile is 24 percent. Hence, for the majority of clusters in our data, the total effect for
private firms is still positive, while that for SOEs and foreign firms is negative, which may be
contrary to what subsidies were intended for.
6. Theoretical interpretation
In order to make sense of these results, we propose a theoretical framework, where we
extend a simple heterogeneous firm type model and show that the treatment effects indeed
depend on the proportion of treated firms, i.e. firms that receive a subsidy. The aim is to give
some theoretical foundation for why both direct and indirect (spillover) effects may depend
on the level of subsidization in a particular cluster. The theoretical predictions on the direct
and indirect productivity effects largely match the empirical results.
We utilize a closed economy heterogeneous firm model a la Melitz (2003), where a
government subsidizes firms by reducing their marginal cost of production. We make the
additional assumption that in clusters with a higher share of subsidized firms, the size of the
per-unit subsidy received by a given firm is lower. 9 The subsidy is financed through a lump
sum tax and is given to firms randomly after they have entered the local market. We choose
this set-up not because we think it is the most realistic – in fact it is well known that
governments are likely to select firms to be subsidized on the basis of observables, as we also
show in our empirical analysis - but because it is most consistent with our empirical
9 In fact, a simple regression of the number of subsidised firms in a cluster on the average subsidy per firm, in
our data returns a negative and statistically significant coefficient. Results are available upon request.
18
identification strategy, which relies on the standard conditional independence assumption of
the treatment evaluation literature.10
There is an exogenous number of L workers, there is no unemployment and workers
supply their time to firms in exchange for a wage, which is normalized to one. Utility of the
individual is represented by a CES utility function
𝑈 ≡ (∫ 𝑑(𝜔)𝛼𝑑𝜔𝑚
0
)
1𝛼
,
where d(ω) denotes demand for product ω and the elasticity of substitution σ = 1/(1-α) > 1 is
determined by the parameter 0 < α < 1. There are infinitely many products ω of mass m, with
m being endogenous. Demand for a product equals
𝑑(𝜔) =𝑝(𝜔)−𝜎
𝑃1−𝜎𝐶,
where p(ω) is the price of the product, C is aggregate consumption expenditure, and P is the
aggregate price index, defined as
𝑃 ≡ (∫ 𝑝(𝜔)1−𝜎𝑑𝜔𝑚
0
)
11−𝜎
.
Firms develop new products, but in order to do that, each firm has to pay a fixed cost
𝐹. The firm then draws marginal cost 𝑎 from a Pareto distribution with a probability density
function 𝑔(𝑎) with support [0, �̅�] and a cumulative density function
𝐺(𝑎) ≡ ∫ 𝑔(𝑎)𝑑𝑎 = (𝑎
�̅�)
𝑘
.𝑎
0
A lump sum tax is given by the government as a subsidy to the marginal cost of
randomly chosen firms from the set of entering firms. Suppose the share of subsidized firms
is 0 < 𝑠 < 1 and the marginal cost of those firms becomes 𝑠𝑎, where 0 < 𝑠 < 1
10 In fact, we use a weaker form of this assumption in that our econometric approach only requires that
conditional on observable firm characteristics and time invariant unobservables (since we use panel data),
subsidy receipt is as good as random. Nonetheless it is fairly straightforward to write down a model with
selection of either the highest or lowest productivity firms receiving subsidies, which would also illustrate the
existence of indirect effects.
19
determines the size of the per-unit subsidy. We assume that 𝑠 increases in 𝑠, with this
assumption the per-unit subsidy received by a given firm decreases in 𝑠. A non-subsidized
firm stays with marginal cost 𝑎.
The effective marginal cost determines how many labor units the firm needs in order
to produce one unit of output. In order to enter the market, each firm also needs to pay a fixed
cost 𝐹𝐿. This setup creates a marginal cost threshold 𝑎𝐿 for entering the local market, which
separates firms that enter from those that are not productive enough and do not enter. The
profit of a non-subsidized firm can be written as 𝜋(𝑎) = (𝑝(𝜔) – 𝑎(𝜔)) 𝑑(𝜔) . Firms
optimize profits and given their marginal cost set the price at 𝑝(𝜔) = 𝑎(𝜔) /𝛼 . Profits
therefore equal 𝜋(𝑎) = 𝛿(𝑎/𝑃)1−𝜎𝐶, where for brevity we write 𝛿 ≡ (𝜎 − 1)𝜎−1 𝜎−𝜎.
Firms face an exogenous exit probability . The government gives a subsidy only to
firms that are able to enter profitably on their own. The expected benefit of entry for the
threshold firm should therefore equal the fixed cost 𝜋(𝑎𝐿)/ = 𝐹𝐿 . In equilibrium the
expected gains from product development, which amount to expected firm value net of the
fixed cost of market entry, have to equal the cost of product development 𝐹. This equality is
expressed in the free entry condition:
𝐹 = 𝑠 ∫ (𝜋(𝑠𝑎)
𝛾− 𝐹𝐿) 𝑔(𝑎)𝑑𝑎
𝑎𝐿
0
+ (1 − 𝑠) ∫ (𝜋(𝑎)
𝛾− 𝐹𝐿) 𝑔(𝑎)𝑑𝑎.
𝑎𝐿
0
We obtain an expression for the marginal cost entry threshold:
𝑎𝐿 = �̅� (𝐹
𝐹𝐿 (𝑆𝑘
𝑘 − 𝜎 + 1− 1)
)
1𝑘
,
where 𝑆 ≡ 𝑠𝑠1−𝜎 + 1 − 𝑠 and
𝜕𝑆
𝜕𝑠= 𝑠
1−𝜎 + (1 − 𝜎)𝑠−𝜎 𝜕𝑠
𝜕𝑠− 1. We have already
stated our assumption that 𝜕𝑠
𝜕𝑠> 0, additionally we assume that for low levels of 𝑠 the
20
derivative 𝜕𝑠
𝜕𝑠 has a high value and decreases in 𝑠, meaning that the second derivative is
negative.
Suppose that there is an 𝑠′ for which in the range 𝑠 ∈ [0, 𝑠′) the derivative 𝜕𝑠
𝜕𝑠 is large
enough so that 𝜕𝑆
𝜕𝑠< 0 and for 𝑠 ∈ [𝑠′, 1] we would have
𝜕𝑆
𝜕𝑠> 0. This implies that in the low
range of 𝑠 between zero and 𝑠′, increasing the share of subsidized firms s would lead to an
increase in the marginal cost threshold 𝑎𝐿 and a decrease in the average firm productivity in
the cluster. More subsidies allow the entry of less productive firms and thus makes the market
less competitive. In the high range of 𝑠 however the opposite happens, the average
productivity in the cluster increases. More details and the solution to the complete model are
provided in the appendix.
We use this model to examine two effects of subsidies, namely, a direct and an
indirect effect. As outlined in Section 4, we define the direct effect as the difference, for a
given cluster with share s, between the average outcome of the subsidized compared to the
non-subsidized firm ( �̅�𝑠𝑠 10 = �̅�𝑠
1 − �̅�𝑠0). The indirect treatment effect, ( �̅�
𝑠0 00 = �̅�𝑠
0 − �̅�00) is the
difference between the outcome of the average non-subsidized firm in a cluster with share of
treated firms s compared to the average in a cluster with no subsidized firms. We look at
productivity of firms to illustrate these effects within the context of our model.
In order to keep the theory as close as possible to the data we will calculate average
productivity of firms based on their effective marginal cost 𝑠𝑎 for subsidized firms and 𝑎
for non-subsidized firms without a productivity improvement.11
The direct effect where productivity is the outcome variable can be written as:
11 We should point out that the productivity measure we employ in the empirical analysis is based on observed firm output. If one were to measure productivity in the same way in our theoretical model, output of a subsidized firm would be higher than the one of a non-subsidized firm with the same marginal cost a. The subsidized firm would, hence, appear to have higher productivity.
21
�̅�𝑠𝑠 10 = ∫ (𝑠𝑎)1−𝜎
𝑔(𝑎)
𝐺(𝑎𝐿)
𝑎𝐿
0
𝑑𝑎 − ∫ 𝑎1−𝜎𝑔(𝑎)
𝐺(𝑎𝐿)
𝑎𝐿
0
𝑑𝑎.
The parameter k of the Pareto distribution has to be larger than σ-1 to make sure that
the above integrals converge. The direct effect is positive since 𝑠1−𝜎 > 1. From the fact that
an increase in the share of subsidized firms in the cluster s first increases the threshold
𝑎𝐿 follows that the average productivity of firms on the market ∫ 𝑎1−𝜎 𝑔(𝑎)
𝐺(𝑎𝐿)
𝑎𝐿
0 will decrease in
s. At the same time 𝑠 increases in s and 𝜅𝑠1−𝜎 decreases. The direct effect is positive and
decreasing in 𝑠. For a large share of subsidized firms in the cluster however the threshold
𝑎𝐿 decreases which leads to an average productivity of firms increasing in s. To summarize:
The direct productivity effect of a subsidy on the subsidy receiving firm is positive and
decreasing in s for 𝑠 ∈ [0, 𝑠′) and positive and increasing in 𝑠 for 𝑠 ∈ [𝑠′, 1].
This result corresponds to our empirical results on the direct productivity effecton
private firms and SOEs to an increase in the share of subsidization in a cluster. Before
moving on, let us spell out the intuition for this a bit more clearly. In the model the non-
monotonic response of the positive direct effect to the increase in the share of subsidized
firms in the cluster is the result of two forces acting in opposite directions. First there is the
assumption that the size of the per-unit-of-production subsidy, and therefore the size of the
subsidy given to a firm, decreases with the share of subsidized firms in the cluster. This
means that a higher share of subsidized firms decreases the size of the positive direct effect.
The lower the subsidy, the lower the positive effect.
Once the subsidy size starts to be less responsive to the share of subsidized firms in a
cluster, a competition effect starts to appear. Subsidizing some entrants changes the
productivity threshold for entry, and thus makes entry of new firms more difficult. New
entrants have to be more productive in a cluster where more firms are subsidized. Hence, this
competition effect implies that the set of producing firms becomes on average more
22
productive given a sufficiently high number of subsidized firms in a cluster (as in Pflüger and
Suedekum, 2013), in turn implying that the direct effect increases in the number of subsidized
firms.12
The indirect effect, where we compare the average productivity of non-treated firms
from a cluster with a share of subsidized firms s versus a cluster without any subsidized firms
(s=0), can be expressed as
�̅�𝑠0 00 = ∫ 𝑎1−𝜎
𝑔(𝑎)
𝐺(𝑎𝐿)
𝑎𝐿
0
𝑑𝑎 − ∫ 𝑎1−𝜎𝑔(𝑎)
𝐺(𝑎𝑛𝐿)
𝑎𝑛𝐿
0
𝑑𝑎.
The threshold 𝑎𝑛𝐿 is the marginal cost threshold in a cluster not receiving any
subsidies. The indirect effect is first negative, because the firms in a subsidized cluster are on
average less productive (aL> anL). Since aL is increasing in s and anL remains constant it is
easy to show that �̅�𝑠0 10 is first decreasing in s. For the high range of 𝑠 we have aL< anL
leading to a positive and increasing indirect effect. The indirect effect in this setup depends
mainly on the presence of the already mentioned competition effect above. To summarize:
The indirect productivity effect is negative and decreasing in s for ∈ [0, 𝑠′) and
positive and increasing in 𝑠 for 𝑠 ∈ [𝑠′, 1].
This matches our empirical results for foreign firms and SOEs in the range of high
and low shares of subsidization. The theory does not match the part with the middle range of
subsidization, where the indirect effect is negative and increasing.
Turning to private firms, in the empirical results the indirect effect for them is a
mirror image to the one for foreign firms and SOEs. A heterogeneous firm model where the
average per-firm subsidy decreases slowly for low values of 𝑠 and then rapidly for high
values of 𝑠 would match this result. A strong competition effect within the low range of s,
12
The competition effect is conditional on the subsidy size not changing too strongly in the share of subsidized
firms. Otherwise the competition effect is reversed and the entry of more firms can also reduce the average
productivity in the cluster.
23
where only more productive firms on average are able to enter a market, would ensure the
initial positive and increasing indirect effect. For a high range of 𝑠 the competition effect
wanes off and on average less productive firms enter a cluster, then the indirect effect would
turn negative and decreasing. Again, the empirically found positive and decreasing indirect
productivity effect would remain unexplained by the theory.
In a standard heterogeneous firm model with subsidies as in Pflüger and Suedekum
(2013) the average firm productivity in a cluster increases as a result of subsidization. In the
data clearly the indirect effect is initially negative for a low share of subsidization. For a
theory to correspond to this result it needs to be based on a model where subsidization leads
in some instances to lower average firm productivity in a cluster. We make in our model the
assumption that the size of the per-unit subsidy decreases with the number of subsidized
firms. There are also other theoretical approaches that can generate this result. In a model
with soft budget constraints for instance a firm’s expectation to receive subsidies may reduce
its managerial effort to maximize profits, reduce costs or invest in innovation, hence leading
to a worsening of firm performance relative to a firm that does not expect to be subsidized
(see Kornai et al., 2003). A similar argument could be made in a model with x-inefficiency
among firms (e.g., Leibenstein and Maital, 1994).
7. Conclusion
In this paper, we implement an approach that allows for the joint estimation of direct
and indirect effects of subsidies on subsidized and non-subsidized firms. In line with much of
the existing literature, we find that firms that receive subsidies experience a boost to
productivity. Looking at it from this angle would then suggest that such a policy “works”.
However, our approach highlights the importance of indirect effects, which are generally not
considered in the literature. We find that, in general but not always, non-subsidized firms
24
experience reductions in their productivity growth if they operate in a cluster where other
firms are subsidized. These negative externalities, and also the positive direct effects, depend
on the share of firms that receive subsidies in a cluster. We interpret our results with a simple
heterogeneous firm type model, which highlights the implications of subsidization for the
competitive environment of firms. Subsidies may potentially harm non-subsidized firms.
Our paper demonstrates the importance of considering indirect effects in evaluation
studies. Not only from a technical perspective (as this improves upon the accuracy of the
results) but, importantly, also from a policy perspective. Taking the effects on non-subsidized
firms into account may significantly change the conclusion as to whether or not the
subsidization policy was beneficial, in terms of improving overall productivity growth in a
local economy. Specifically, our findings, contrasting with advocates of state capitalism,
provide empirical evidence that highlights the potential cost of state intervention through
subsidization, in terms of negative externalities. Overall, our estimation approach allows us to
provide a much richer analysis on the relationship between subsidies and firm performance
than the literature thus far.
25
References
Aghion, P., J.Cai, M. Dewatripont, L. Du, A. Harrison, P. Legros (2015), Industrial Policy
and Competition, American Economic Journal: Macroeconomics, 7(4): 1-32
Ackerberg, D. A., Caves, K. and Frazer, G. (2015), Identification Properties of Recent
Production Function Estimators. Econometrica, 83: 2411–2451.
Bang, H. and Robins, J. M. (2005). Doubly robust estimation in missing data and causal
inference models, Biometrics 61,962-972.
Bellak, C., (2004). How domestic and foreign firms differ and why does it matter? Journal of
Economic Surveys, 18(4), 483-514.
Bernini, Cristina & Pellegrini, Guido (2011). How are growth and productivity in private
firms affected by public subsidy? Evidence from a regional policy, Regional Science and
Urban Economics, Elsevier, 41(3), 253-265.
Branstetter, L.G. and Feenstra, R.C. (2002), Trade and foreign direct investment in China: a
political economy approach, Journal of International Economics 58, 335-358.
Chang, S.J. and S.Y. Jin, 2016, The performance of state-owned enterprises in China: An
empirical analysis of ownership control through SASACs, Centre for Governance, Institutions
and Organisations (CGIO), NUS Business School National University of Singapore
(https://bschool.nus.edu.sg/Portals/0/docs/CGIO/soe-china-research-report-2016.pdf)
Cerqua, A and Pellegrini, G (2014). "Do subsidies to private capital boost firms' growth? A
multiple regression discontinuity design approach," Journal of Public Economics, 109(C)
114-126.
De Mel, S., McKenzie, D., & Woodruff, C. (2008). Returns to capital in microenterprises:
evidence from a field experiment. Quarterly Journal of Economics, 1329-1372
Du, J., and Girma, S. (2010). “Red capitalists: Political connections and firm performance in
China”, Kyklos, 63(4), 530-545.
Du, J., X. Liu, and Y. Zhou (2014). "State advances and private retreats?—Evidence of
aggregate productivity decomposition in China." China Economic Review 31: 459-474.
26
Du, & Mickiewicz, T. (2016). Subsidies, rent seeking and performance: Being young, small
or private in China. Journal of Business Venturing, 31(1), 22-38.
Girma, S, Gong, Y; Görg, H Yu, Z (2009). "Can Production Subsidies Explain China's
Export Performance? Evidence from Firm-level Data," Scandinavian Journal of Economics,
111(4), 863-891.
Girma, S., Gong, Y., Görg, H. and Lancheros, S. (2015). “Estimating direct and indirect
effects of foreign direct investment on firm productivity in the presence of interactions
between firms”, Journal of International Economics, 95, 157–169.
Görg, H., Henry, M., Strobl, E. (2008). Grant support and exporting activity. Review of
Economics and Statistic, 90(1), 168-174
Haley, U. and G. Haley (2013a). Subsidies to Chinese Industry. Oxford, UK: Oxford
University Press.
Haley, U. and G. Haley (2013b). “How Chinese subsidies changed the world”, Harvard
Business Review, April 2013, available at https://hbr.org/2013/04/how-chinese-subsidies-
changed
Hirano, K. and Imbens, G.W. (2004). `The propensity score with continuous treatments’, in:
Andrew Gelman and Xiao-Li Meng eds, Applied Bayesian Modeling and Causal Inference
from Incomplete-Data Perspectives, Wiley.
Howell, A. (2017), Picking ‘winners’ in China: Do subsidies matter for indigenous
innovation and firm productivity?, China Economic Review, 44, 154-165
Hsieh, C. and Klenow, P., (2009), Misallocation and manufacturing TFP in China and India,
Quarterly Journal of Economics, Vol. 124, pp.1403-1448.
Hudgens, M.G., and Halloran, M.E. (2008). “Toward Causal Inference with Interference’,
Journal of the American Statistical Association, 103, 832-842.
Imai K. and van Dyk D.A. (2004). “Causal inference with general treatment regimes:
Generalizing the propensity score’, Journal of the American Statistical Association 99, 854-
866.
27
Klitgaard, T.J. and Rasmussen, M., 1983. Preferential Treatment for Foreign Investment in
the People's Republic of China: Special Economic Zones and Industrial Development
Districts. Hastings Int'l & Comp. L. Rev., 7, p.377.
Kornai, J., E. Maskin and G. Roland (2003), “Understanding the soft budget constraint”,
Journal of Economic Literature, 41, 1095-1136.
Leibenstein, H. and S. Maital, 1994, The organizational foundations of X-inefficiency: A
game-theoretic interpretation of Argyris' model of organizational learning, Journal of
Economic Behavior & Organization, 23(3), 251-268
Lin, Yifu J. and Li, Zhiyun D. (2008). “Policy Burden, Privatization and the Soft Budget
Constraint”, Journal of Comparative Economics 36(1), 90-102
Luo, Huaping & T. Golembiewski, Robert. (1996). Budget deficits in China: Calculations,
causes, and impacts. Journal of Public Budgeting, Accounting & Financial Management. 1.
32-54.
Melitz, M. (2003). “The impact of trade on intra‐industry reallocations and aggregate
industry productivity,” Econometrica, 71(6), 1695-1725.
Mueller, D.C. 2003. Public Choice III. Cambridge: Cambridge University Press.
Papke, .E. and Wooldrige, J. (1996). “Econometric methods for fractional response variables
with an application to 401(k) plan participation rates”, Journal of Applied Econometrics, 11,
619-632.
Pflüger, M. and J. Suedekum (2013). Subsidizing firm entry in open economies, Journal of
Public Economics, 97, 258-271
Shao Min and Bao Qun, (2011). “Analysis of government subsidy: support the strong or
protect the weak?” World Economic Papers, No.1, February, page 56-72 (in Chinese)
Xu, Chenggang, (2011). “The Fundamental Institutions of China’s Reforms and
Development”, Journal of Economic Literature, 49:4, 1076–1151
28
Woo, Wing Thye, Lu, Ming, Sachs, Jeffrey D., 2012. A New Economic Growth Engine for
China: Escaping the Middle-income Trap by Not Doing More of the Same. World Scientific
Publishing Company.
World Trade Organization, (2006) World Trade Report - Exploring the links between
subsidies, trade and the WTO, WTO
29
FIGURES
Figure 1
30
Figure 2
31
Figure 3:
32
Figure 4:
33
TABLES
Table 1: Definition of variables used in the analysis
Variables Definition
Pre-treatment covariates
Past Subsidy Dummy variable indication whether the firm received subsidy in the past two
years
Employment Log of number of employees employment
Age Log of firm age
Loss Dummy variable indicating whether the firms was a loss making one in the past
two years
TFP Total factor productivity calculated using Ackerberg, Caves and Frazer (2015)
correction, with the amount of subsidies included in the production function as a
state variable.
TFP trend Firms TFP growth relative to industry median trend
Debt Total liability of total assets
POL Dummy variable indicating whether the firm enjoys political affiliation with local
and central governments
Spatial variables All defined as spatial averages within the town the firm is located in (the
cluster) but excluding one’s own value. The prefix S_ is used to indicate
spatial averages.
S_Past Subsidy; S_Employment; S_Age; S_Loss; S_TFP; S_TFP trend; S_Debt; S_POL
Dummy variables with private firms, low tech used as base groups.
Foreign Dummy variable indicating whether a firm is foreign owned
SOE Dummy variable indicating whether a firm is a state-owned enterprise
Medium low tech Dummy variable indicating whether a firm operates in a medium low tech
industry, using OECD classification scheme, see
http://www.oecd.org/sti/ind/48350231.pdf.
Medium high tech Dummy variable indicating whether a firm operates in a medium high tech
industry.
High tech Dummy variable indicating whether a firm operates in a high tech industry.
Outcome variable
TFP change TFP change after receipt of subsidy relative to the pre-treatment period, Baseline
regressions conducted with respect to changes one year after the treatment period.
This approaches allows for a difference-in-differences type approach to control for
time-invariant unobservables.
Treatment variables
Subsidy Dummy variable indication whether the firm received subsidy in the current period
Indirect effects
capturing variable
S_subsidy Proportion of subsidized firms (excluding one’ own value) in the firm’s town.
34
Table 2:
Total subsidies in Billions of USD and the number of
of subsidized firms by ownership
Private firms Foreign firms SOEs
Year SUBSIDY # of
firms
SUBSIDY # of
firms
SUBSIDY # of
firms
1998 3.618 5897 1.070 1369 6.118 6588
1999 4.482 6244 1.099 1805 6.311 6514
2000 5.666 7509 1.093 2038 6.291 5878
2001 7.314 8654 1.458 2836 6.118 5384
2002 9.562 10751 1.896 4064 6.458 5259
2003 11.617 13364 2.541 5419 6.250 5015
2004 17.073 22229 3.698 9421 6.671 5457
2005 17.413 21049 3.892 7354 6.779 4833
2006 20.776 22994 4.815 8240 6.952 4635
2007 23.662 25673 5.887 8565 6.727 4077
35
Table 3: Summary statistics by subsidy status
Non-Subsidized Subsidized
(N=800103) (N=144655)
Mean Std. dev Mean Std. dev
TFP growth 0.014 0.423 0.006 0.359
Past Subsidy 0.064 0.245 0.619 0.486
Employment 4.880 1.077 5.312 1.186
Age 2.187 0.772 2.272 0.800
Loss 0.209 0.406 0.279 0.448
TFP 0.826 0.472 0.834 0.480
TFP trend 1.228 184.465 1.393 121.249
Debt 0.057 0.140 0.060 0.125
POL 0.478 0.500 0.528 0.499
S_Past subsidy 0.127 0.069 0.168 0.077
S_Employment 4.827 0.247 4.780 0.234
S_Age 1.993 0.212 1.984 0.196
S_Loss 0.218 0.098 0.215 0.103
S_TFP 0.828 0.107 0.826 0.103
S_TFP trend 1.180 6.472 1.098 6.180
S_Debt 0.056 0.032 0.051 0.031
S_POL 0.461 0.284 0.438 0.279
Private firms 0.609 0.488 0.561 0.496
Foreign firms 0.221 0.415 0.225 0.418
SOEs 0.170 0.376 0.214 0.410
Low-tech 0.362 0.481 0.314 0.464
Medium low-tech 0.325 0.468 0.332 0.471
Medium tech 0.215 0.411 0.243 0.429
High-tech 0.098 0.297 0.111 0.314
36
Table 4: The conditional probability of subsidy:
log odds ratios from spatial logistic regression: Log odds ratio Robust standard
errors
Past Subsidy 2.915*** (0.00914)
Employment 0.287*** (0.00370)
Age -0.0542*** (0.00531)
Loss making 0.141*** (0.00889)
TFP 0.0639*** (0.00828)
TFP growth trend 0.0000159 (0.0000185)
Debt 0.0603* (0.0268)
Political connection 0.178*** (0.00937)
S_past Subsidy 4.656*** (0.0201)
S_employment -0.303*** (0.0296)
S_Age -0.110*** (0.0623)
S_Loss making -0.218*** (0.0609)
S_TFP -0.847*** (0.000510)
S_TFP growth trend -0.000943 (0.238)
S_Debt -1.949*** (0.0328)
S_ POL -0.268***
FOREIGN -0.0127 (0.00973)
SOE 0.112*** (0.0117)
Medium low-tech
industries
0.183*** (0.00931)
Medium high-tech
industries
0.147*** (0.0105)
High-tech industries 0.196*** (0.0130)
Observations 944758
Log likelihood -283030.6
Pseudo R-squared 0.300
Notes:
(i) Private firms and low-tech industries form their respective base group.
(ii) All regressions include year effects.
(iii) * p<0.1, ** p<0.05, *** p<0.01
37
Appendix
Table A1: The conditional probability of subsidy:
log odds ratios from spatial logistic regression: by ownership Private Foreign SOEs
Log odds ratio
Robust s/e
Log odds ratio
Robust s/e
Log odds ratio
Robust s/e
TFP growth 3.090*** (0.0124) 2.369*** (0.0177) 2.974*** (0.0208)
Past Subsidy 0.289*** (0.00539) 0.242*** (0.00750) 0.316*** (0.00720)
Employment -0.0492*** (0.00706) -0.155*** (0.0148) -0.0457*** (0.0102)
Age 0.237*** (0.0128) -0.130*** (0.0183) 0.112*** (0.0172)
Loss 0.122*** (0.0118) -0.0363* (0.0171) 0.0631*** (0.0157)
TFP 0.0000130 (0.0000278) -0.0000293 (0.0000411) 0.0000495 (0.0000298)
TFP trend -0.0351 (0.0391) 0.265*** (0.0737) 0.0894 (0.0469)
Debt 0.236*** (0.0120) 0.0173 (0.0192) 0.0228 (0.0279)
POL 5.093*** (0.0803) 4.854*** (0.131) 2.730*** (0.128)
S_Past subsidy -0.222*** (0.0299) -0.538*** (0.0498) -0.0857* (0.0359)
S_Employment -0.187*** (0.0402) -0.111 (0.0789) -0.132* (0.0596)
S_Age 0.209* (0.0872) -0.690*** (0.161) 0.252* (0.124)
S_Loss -0.675*** (0.0851) -1.894*** (0.165) -0.164 (0.118)
S_TFP -0.00200 (0.00103) 0.00332 (0.00209) -0.000873 (0.000633)
S_TFP trend -4.224*** (0.352) 0.543 (0.741) -0.116 (0.387)
S_Debt -0.0994* (0.0444) -0.344*** (0.0756) -0.436*** (0.0769)
S_POL 0.211*** (0.0125) 0.0938*** (0.0196) 0.160*** (0.0214)
Medium high-tech 0.151*** (0.0142) 0.153*** (0.0214) 0.112*** (0.0242)
High-tech 0.211*** (0.0181) 0.182*** (0.0256) 0.139*** (0.0280)
Observations 568323 209071 167364
Log likelihood -157174.7 -69499.5 -54360.3
Pseudo R-squared 0.325 0.232 0.322
Notes:
(i) Low tech industries used as the base group
(ii) * p<0.1, ** p<0.05, *** p<0.01
38
Table A2:
Covariate balance tests based on standardized differences
and variance ratios: Private firms
Standardised
differences
Variance
ratio
Raw
data
Weighted
data
Raw
data
Weighted
Data
Pre-treatment
covariates
TFP growth 152.24% 0.85% 4.452 1.017
Past Subsidy 33.12% 4.31% 1.260 1.096
Employment 11.54% -1.64% 0.985 0.971
Age 23.72% 2.36% 1.465 1.044
Loss 3.17% -0.48% 1.055 1.007
TFP -0.03% 0.26% 0.446 0.507
TFP trend -1.16% 0.25% 0.779 0.915
Debt 12.07% -0.89% 1.030 0.997
POL 64.53% 8.38% 1.151 0.888
S_Past subsidy -22.38% -0.69% 0.882 0.995
S_Employment -4.00% -0.17% 0.747 0.957
S_Age -3.54% 2.27% 1.077 1.034
S_Loss 1.23% -2.23% 0.893 1.060
S_TFP -3.11% -0.72% 0.950 1.081
S_TFP trend -23.97% -0.55% 0.816 1.021
S_Debt -9.82% -0.33% 0.954 0.984
S_POL 2.30% -2.85% 1.013 0.982
Medium high tech 6.99% 0.41% 1.096 1.006
High tech 4.52% 2.06% 1.133 1.059
Year=2001 -5.73% 1.81% 0.813 1.063
Year=2002 -2.36% -1.99% 0.925 0.936
Year=2003 -1.04% -2.05% 0.973 0.946
Year=2004 10.87% 0.88% 1.322 1.025
Year=2005 6.20% 1.72% 1.146 1.039
Year=2006 -1.37% -0.29% 0.982 0.996
Year=2007 -1.97% -0.87% 0.976 0.990
39
Table A3:
Covariate balance tests based on standardized differences
and variance ratios: Foreign firms
Standardised
differences
Variance
ratio
Raw
data
Weighted
data
Raw
data
Weighted
Data
Pre-treatment
covariates
TFP growth 115.557% 0.841% 3.388 1.016
Past Subsidy 23.000% -0.802% 0.990 0.970
Employment -5.869% -0.730% 1.093 1.066
Age -10.693% 0.762% 0.859 1.010
Loss -0.851% 0.976% 0.951 1.140
TFP 0.006% -0.199% 0.405 0.411
TFP trend 2.254% 0.967% 0.658 0.585
Debt -4.601% -1.536% 0.968 0.990
POL 63.286% 4.081% 1.229 0.919
S_Past subsidy -41.155% -3.164% 0.727 0.895
S_Employment -11.094% -1.543% 0.888 1.029
S_Age -27.415% 0.817% 1.048 0.999
S_Loss -6.079% 1.159% 0.823 1.043
S_TFP 3.238% 1.299% 1.085 0.977
S_TFP trend -21.087% -1.571% 0.879 1.000
S_Debt -13.146% -1.105% 0.936 0.992
S_POL -4.967% -1.331% 0.943 0.985
Medium high tech 5.920% 0.244% 1.076 1.003
High tech 3.397% -0.409% 1.090 0.989
Year=2001 -10.187% -1.848% 0.707 0.943
Year=2002 -2.827% -1.756% 0.919 0.949
Year=2003 3.334% 1.284% 1.088 1.033
Year=2004 16.915% -0.208% 1.463 0.995
Year=2005 1.551% 6.454% 1.034 1.142
Year=2006 1.253% -0.520% 1.019 0.992
Year=2007 -1.212% -3.458% 0.983 0.950
40
Table A4:
Covariate balance tests based on standardized differences
and variance ratios: SOEs
Standardised
differences
Variance
ratio
Raw
data
Weighted
data
Raw
data
Weighted
Data
Pre-treatment
covariates
TFP growth 155.445% 0.948% 3.117 1.015
Past Subsidy 58.520% 3.910% 1.034 1.072
Employment 7.462% -3.632% 1.094 1.058
Age 15.172% -1.225% 1.033 0.996
Loss 4.474% 0.786% 0.989 1.035
TFP 0.313% 0.081% 0.380 0.494
TFP trend 4.207% -0.636% 0.771 0.829
Debt -0.094% -2.699% 1.002 1.070
POL 34.659% 4.343% 1.446 0.890
S_Past subsidy -1.931% 3.267% 0.912 0.965
S_Employment -15.184% -0.496% 0.998 1.039
S_Age -3.948% 1.803% 1.107 1.055
S_Loss 7.578% -1.129% 0.988 1.020
S_TFP -1.370% 0.497% 0.731 0.661
S_TFP trend -11.115% 0.927% 0.960 1.038
S_Debt -19.083% 0.096% 1.063 1.009
S_POL 9.675% -0.449% 1.079 0.996
Medium high tech 5.997% -0.162% 1.079 0.998
High tech 2.377% 0.490% 1.055 1.011
Year=2001 -9.103% 0.096% 0.833 1.002
Year=2002 -4.494% -1.202% 0.909 0.975
Year=2003 -1.389% -0.638% 0.970 0.986
Year=2004 7.074% -0.044% 1.199 0.999
Year=2005 8.228% -0.937% 1.243 0.974
Year=2006 8.643% -0.632% 1.239 0.983
Year=2007 8.269% 1.614% 1.252 1.046
41
Table A5:
Direct, indirect and (weighted) total treatment effects
Private firms
Direct effects Indirect effects Weighted total effects
% of subsidized
0,05 0,018 0,014 0,014
0,10 0,010 0,021 0,020
0,15 0,007 0,020 0,018
0,20 0,007 0,013 0,012
0,25 0,011 0,001 0,004
0,30 0,018 -0,016 -0,006
0,35 0,027 -0,036 -0,014
0,40 0,039 -0,060 -0,020
0,45 0,052 -0,087 -0,024
0,50 0,067 -0,116 -0,025
SOEs
Direct effects Indirect effects weighted total effects
% of subsidized
0,05 0,018 -0,017 -0,015
0,10 0,009 -0,030 -0,026
0,15 0,005 -0,041 -0,034
0,20 0,006 -0,049 -0,038
0,25 0,011 -0,056 -0,039
0,30 0,019 -0,060 -0,036
0,35 0,031 -0,063 -0,030
0,40 0,045 -0,065 -0,021
0,45 0,061 -0,064 -0,008
0,50 0,079 -0,062 0,009
Foreign firms
Direct effects Indirect effects Weighted total effects
% of subsidized
0,05 0,012 -0,031 -0,029
0,10 0,004 -0,052 -0,046
0,15 0,008 -0,065 -0,054
0,20 0,016 -0,070 -0,053
0,25 0,024 -0,068 -0,045
0,30 0,029 -0,060 -0,033
0,35 0,032 -0,045 -0,018
0,40 0,034 -0,026 -0,002
0,45 0,035 -0,002 0,015
0,50 0,035 0,026 0,031
Recommended